Deep asymptotic expansion method for solving singularly perturbed time-dependent reaction-advection-diffusion equations

Bangti Jin; Dmitrii Chaikovskii; Qiao Zhu; Ye Zhang

REVIEW 3 major objections 5 minor 38 references

Deep asymptotic expansion method for solving singularly perturbed time-dependent reaction-advection-diffusion equations

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Learn the front position, not the whole solution field.

desk verdict The DAE method is a sensible new combination of asymptotics and PINN, but its numerical validation is circular and the outperformance claim is not yet established. read the letter →

arxiv 2505.23002 v2 pith:2EVS4EU2 submitted 2025-05-29 math.NA cs.NA

Qiao Zhu , Dmitrii Chaikovskii , Bangti Jin , Ye Zhang This is my paper

classification math.NAcs.NA MSC 35B2565D1765N9968T07

keywords singularlyperturbedPDEreaction-advection-diffusionequationasymptoticexpansionphysics-informedneuralnetworkinternaltransitionlayerdeepadaptivesamplingmovingfront

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Solutions of singularly perturbed reaction-advection-diffusion equations are smooth almost everywhere but develop a steep internal transition layer of width $O(\mu|\ln\mu|)$. This paper claims that for such problems the right learning target is not the full solution but the position $h_0(x_*,t)$ of the front. The authors derive an integral equation that $h_0$ must satisfy, train a small physics-informed neural network to solve that reduced equation, and assemble a zeroth-order asymptotic solution from it. They prove that the assembled solution satisfies $|u-U_0|\le C_1\mu+C_2(\varepsilon_{\mathrm{app}}+\varepsilon_{\mathrm{gen}}+\varepsilon_{\mathrm{opt}})$, and in one, two and three dimensions they report relative $L^2$ errors near $10^{-6}$ to $10^{-2}$, orders of magnitude below standard PINN, gPINN, and PINN with adaptive sampling, in less training time.

What carries the argument

The load-bearing device is the matched asymptotic expansion: the solution is split into outer functions $\varphi^{(\mp)}(x)$ and inner transition-layer functions $Q_0^{(\mp)}(\xi,h_0,x_*,t)$ in the rescaled normal coordinate $\xi=r_1/\mu$. The front position $h_0(x_*,t)$ is determined by the integral equation (2.17), which balances the integral of the advection-speed terms between the two outer branches. DAE parameterizes $h_0$ by a network of the form $h_0^*+t\,\mathcal{N}(x_*,t;\tilde\theta)$ so that the initial front position is exact by construction, and minimizes a loss consisting of Gauss-Legendre quadrature residuals of (2.17) plus a periodicity penalty. RAR then reweights the collocation set toward points where the reduced-equation residual is largest. The assembled solution from (2.20) inherits accuracy from both the analytic asymptotic structure and the trained front.

What would settle it

Run a 1D problem with initial data containing two well-separated tanh transition layers so that the front position is not single-valued; the DAE ansatz with one $h_0$ will fail and the relative $L^2$ error should jump by orders of magnitude relative to the single-front case, showing that the single-front assumption is load-bearing.

Watch

Extended reading notes

Core claim

The central claim is that the matched asymptotic structure of the problem reduces the singular-perturbation problem to a lower-dimensional, smoother problem that neural networks solve well. The solution is written as $U_0=\varphi^{(\mp)}(x)+Q_0^{(\mp)}(\xi_0,h_0,x_*,t)$, with $\xi_0$ the rescaled normal distance to the front; Theorem 2.4 states that under Assumption 1 this gives $|u-U_0|=O(\mu)$ and $|h-h_0|=O(\mu|\ln\mu|)$. Theorem 3.2 extends this to the trained DAE approximation, showing the total error is $C_1\mu$ plus the usual approximation, generalization, and optimization errors of the network that learns $h_0$. The numerical section demonstrates that this reduction is not just theoretical: DAE and DAE+RAR reach relative $L^2$ errors of $10^{-6}$ to $10^{-2}$ in 1D, 2D, and 3D, whereas PINN, gPINN, and PINN+RAR stall around $0.1$ to $1$, and the DAE trains in roughly a quarter to a half of the time.

Load-bearing premise

The construction assumes the true solution has exactly one internal transition layer already formed at the initial time, and that explicit formulas for the outer functions $\varphi^{(\mp)}$ and the zero-order inner functions $Q_0^{(\mp)}$ are available; when those formulas are missing, the clean error bound no longer applies directly.

Editorial extensions

If this is right

The PINN spectral-bias failure on steep layers is bypassed because the network never sees the layer; it sees only a smooth front-position field.
As $\mu\to0$, the leading $C_1\mu$ term in the error bound shrinks, so the attainable accuracy is governed by how well the front is learned, not by the layer width.
The same reduction is dimension-adaptive: the layer-position equation lives in the $d-1$ transverse variables, so 2D and 3D problems reduce to 1D and 2D learning tasks, respectively.
With RAR, the adaptive points concentrate where the front has high curvature or rapid temporal change, and the reported errors drop by roughly two to three orders of magnitude compared with DAE without RAR.
First-order corrections give only marginal accuracy gains over the zeroth-order DAE while more than doubling training time, so the paper argues zeroth-order is the practical choice.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The method's practical reach is probably the linear-advection case $A=-u$ plus a narrow class with explicit $\varphi^{(\mp)}$ and $Q_0^{(\mp)}$; for general nonlinear $A$ the paper itself notes the inner and outer functions must be network-approximated, which could reintroduce the stiffness it avoids. A direct test would be to run DAE on $A=u^2$ or $A=\sin u$ against a fine-mesh reference.
The single-front ansatz suggests a failure mode the paper does not address: if two fronts merge, split, or are born after $t=0$, no single-valued $h_0$ exists and the reduction breaks; checking the DAE on such data would delimit the method's validity.
The same 'learn the interface, assemble the solution' template might carry over to boundary layers, shocks, and other singularly perturbed PDEs, but each application requires its own asymptotic reduction, so the template is a design principle rather than a drop-in algorithm.
The reported robustness to seeds, depth, width, and sampling distributions hints, but does not prove, that the reduced loss landscape is much better conditioned than the full PINN loss; analysing that conditioning would be a natural theoretical follow-up.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First, the stress-test note is right and it lands. The paper's reference solution is U0, the same zeroth-order asymptotic expansion the DAE reconstructs. The reported relative L2 errors of ~1e-6 at mu=1e-2 are far below the O(mu) bound Theorem 2.4 gives for the actual PDE solution, which is only possible because the baseline is U0. So the comparisons in Tables 2-11 score PINN against the full stiff PDE and score DAE against its own low-dimensional reduction. That is not apples-to-apples.

What is genuinely new here is training a PINN on the residual of the integral equation (2.17) for the front position, rather than on the full PDE. That is a legitimate and potentially useful idea. Embedding h0(0) in the network and adding residual-based adaptive sampling are sensible choices. The paper ships code and data, and it shows robustness across seeds and architectures. The asymptotic derivation is also extended to a general nonlinear A(u,x), but the testbed never leaves A=-u and the proof of Theorem 2.4 sits in the supplement, which I did not check.

Soft spots, in proportion. The circular reference is the main one, and it is not a minor artifact: it means the reported e2 and e_infinity numbers are fit-to-ansatz errors, not errors against a solution of (1.1). Relatedly, Theorem 3.2 leaves the O(mu) term in the bound as an unquantified constant, and no experiment measures it. The generality claim is also untested—Remark 3.1 explicitly concedes the method needs closed-form outer and inner functions, and every experiment uses A=-u.

None of this kills the contribution. For the case the paper actually solves, the approach is cheap, stable, and faster than direct PINN, and the front-equation-fitting part is demonstrated convincingly. What is not demonstrated is that the assembled asymptotic solution solves (1.1) to the claimed accuracy.

I would send this to peer review, with a required addition: at least one independent comparison against a resolved reference solution of the original PDE for A=-u, ideally on a manufactured solution or a fine-mesh small-mu case where the true solution is computable. At that point the reader can separate the method's own asymptotic error from the front-network fitting error. Without that, the accuracy tables should be read as evidence about the network fitting h0, not about solving the PDE.

The audience is people working at the asymptotics-ML interface for singularly perturbed problems. I would not take the error tables as independent evidence, but I would cite the method as a sensible hybrid and would want to see the revised version.

Referee Report

3 major / 5 minor

Summary. The paper proposes a 'deep asymptotic expansion' (DAE) method for singularly perturbed time-dependent reaction-advection-diffusion equations with internal transition layers. The authors derive an asymptotic expansion that separates the solution into outer and inner parts, reduce the problem to an integral equation for the transition-layer position h0(x*,t), and solve this reduced equation with a physics-informed neural network (PINN). The method is tested on 1D, 2D, and 3D problems and compared against standard PINN, gradient-enhanced PINN, and PINN with residual-based adaptive refinement. The paper reports significantly higher accuracy and lower training time for DAE in all cases.

Significance. The asymptotic derivation is clearly presented for the linear advection case, and the idea of using a neural network to solve the reduced front equation is conceptually appealing. The paper makes code and data publicly available, and the robustness studies across random seeds, network sizes, and training data sizes are a useful addition. The error decomposition in Theorem 3.2 provides a formal framework for separating approximation, generalization, and optimization errors. However, the main numerical claim that DAE outperforms PINN on the original PDE is not supported by the current validation, which compares against the asymptotic reference U0 rather than the true solution u. If the authors add an independent reference solution and quantify the O(μ) asymptotic error, the method could be a valuable contribution; as it stands, the evidence is circular for the central claim.

major comments (3)

[Section 4 (error metrics) and Tables 2, 6, 10] The error metrics e2 and e∞ defined at the beginning of Section 4 are computed against 'the reference solution mentioned above', which is the zeroth-order asymptotic solution U0 from Eq. (2.20), not the true solution u of Eq. (1.1). Since the DAE is constructed to reproduce U0 by fitting h0 to Eq. (2.17), the reported e2/e∞ values measure the network's fit to the asymptotic ansatz. The comparison with PINN is therefore not apples-to-apples: PINN is trained on the residual of the full PDE (1.1) but scored against U0, so its error includes the O(μ) asymptotic gap |u-U0|, while DAE's error excludes that gap by construction. For instance, Table 2 reports DAE+RAR e2=2.92e-06 at μ=1e-2, yet Theorem 2.4 guarantees only |u-U0|=O(μ)≈1e-2; such small errors are possible only because the baseline is U0 itself. The paper should validate against an independent numerical solution of (1.1) or at least report the estimated asymptotic error |u-U0| and decompose the total error.
[Theorem 3.2 (Eq. (3.5))] The bound in Eq. (3.5) contains the term C1 μ, where C1 is an unspecified constant. None of the experiments attempt to estimate C1 or to measure |u(x,t)-U0(x,t)| directly. Consequently, the reported accuracies on the order of 1e-6 to 1e-2 cannot be interpreted as accuracies with respect to u; the theorem only provides an O(μ) floor that is not quantified. To support the central accuracy claim, the paper needs either an estimate of C1 for the test problems or a numerical comparison against a high-resolution reference solution of the full PDE.
[Sections 4.1–4.3 and Remark 3.1] All numerical experiments use A(u,x) = -u, the linear case for which explicit formulas for φ(±) and Q0^(∓) are given in Remark 2.3. Remark 3.1 concedes that the DAE requires explicit formulas for these functions, and that when they are unavailable they must themselves be approximated by neural networks. The paper's stated contribution of handling 'a general nonlinear advection coefficient' (Section 1) is therefore not supported by any numerical evidence. Either a nonlinear example with explicit asymptotics (e.g., a cubic A) should be included, or the generality claims should be explicitly restricted to the linear case.

minor comments (5)

[Section 4.1] The statement 'Table 2 indicates that RAR does not improve the accuracy of PINN' is contradicted by Table 2, where PINN+RAR reduces e2 from 7.79e-01 to 1.63e-01 at μ=1e-2; the text should be corrected to say that RAR improves PINN but not to the level of DAE.
[Eq. (3.2)] The formula for the residual R_j is typeset ambiguously; the denominator sqrt(1+Σ(∂xi h0)^2) should be applied to the entire integrand as in Eq. (2.17). Please clarify the formula.
[Section 4.1–4.3] The verification of Assumption 1 is incomplete: Section 4.1 states that Assumptions 1(a), (b), and (d) are satisfied without demonstration, and Sections 4.2–4.3 do not explicitly verify Assumption 1(e). Since Theorem 2.4 depends on Assumption 1, explicit verification for each example would strengthen the validity of the asymptotic reference.
[Eq. (3.1)] The network ansatz ĥ0 = h*_0 + t·h(x*, t; θ) is only valid for a constant initial front h*_0; the paper should state this restriction or generalize the ansatz for x*-dependent initial fronts.
[Tables 2, 6, and 10] The note about dashes is ambiguous: the dashes apply only to e_loss and Time, while e2 and e∞ are new values. Consider clarifying the notation in the table captions.

Circularity Check

1 steps flagged · score 6.0 of 10

Validation is self-referential: reported errors are measured against the DAE's own asymptotic-ansatz reference, not against the PDE solution.

fitted input called prediction [Section 4 (Experimental results), error definitions; Algorithm 3.1 steps 10–13; Eq. (2.20); Theorem 2.4]
"We employ the problems and the asymptotic solutions from existing works [10, 9, 11] for the evaluation. ... where u(xi, ti) represents the reference solution mentioned above."

Algorithm 3.1 constructs U0(x,t;theta*) by assembling Eq. (2.20) with the trained network h0 in place of the exact h0. The 'reference solution' u(xi,ti) used in the reported errors is the same zeroth-order asymptotic solution U0(x,t) from Eq. (2.20) with the exact h0. Thus e2 and e_inf measure only how well the network fits h0 to Eq. (2.17); they do not measure error against the true PDE solution u. Theorem 2.4 states |u-U0|=O(mu); at mu=1e-2 the reported DAE+RAR e2=2.92e-06 is far below any O(mu) gap, which is possible only because the baseline is U0 itself. The comparison is therefore not apple-to-apples: PINN is scored against U0 (including the O(mu) model error) while DAE is scored against its own ansatz.

full rationale

The asymptotic reduction in Section 2, from the PDE (1.1) to the layer-position equation (2.17) and the representation (2.20), is a legitimate mathematical derivation and is not circular in itself. The circularity enters at the evaluation stage. Algorithm 3.1 returns U0(x,t;theta*) built from Eq. (2.20) with a trained h0, while the 'reference solution' used in the error metrics is the same Eq. (2.20) with the exact h0, taken from prior works by the same research group [9,10,11]. Consequently the vanishingly small reported errors (e.g., 2.92e-06 at mu=1e-2) measure the network's fit to h0, not the distance from the true PDE solution u. Theorem 3.2's bound |u-U0(·;theta*)| <= C1 mu + C2(...) is not tested because the C1 mu term is never evaluated experimentally. The paper's genuine contribution—solving the front-position equation with a PINN—remains, but the central advertised claim that DAE 'outperforms' PINN for the original problem is not established by the numerical validation as presented. The self-consistency of the test makes the numerical evaluation partially circular, while the underlying asymptotic derivation itself is not definitionally circular.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The derivation rests on matched asymptotic expansion assumptions and on explicit formulas for the outer and inner functions. The general nonlinear case advertised in the title is not realized numerically; all experiments use A=-u. No independent numerical reference to the full PDE is used, so the empirical accuracy is to a large extent self-referential.

free parameters (2)

Trained neural network weights and biases for h0 (theta) = Not reported as a fixed vector; optimized by Adam
The DAE output U0 depends on the learned parameters theta*, and Theorem 3.2 bounds the resulting error only by the abstract sum of approximation, generalization, and optimization errors, which is never quantified in the experiments.
Training and RAR hyperparameters = Depth 3-6, width 8-16, iterations 8000-40000, RAR batch sizes and tolerance 1e-6; see Table 1 and Section 4
Chosen by hand per problem; robustness is checked only for architecture size, training data size, and random seeds, not for these settings.

assumptions (5)

domain assumption Assumption 1(a)-(e), including existence of outer solutions phi(+/-), sign conditions on A, a formed front at t=0, and the integral positivity condition in (e)
Invoked before Theorem 2.4 and again in Section 4 to guarantee existence, uniqueness, and monotonicity of the transition layer solution.
domain assumption Matched asymptotic ansatz (2.6)-(2.8): U_n = outer part + inner transition layer Q, with Q decaying exponentially
This is the central decomposition of the solution; no proof is given that every solution of (1.1) has this form, conditional only on Assumption 1.
domain assumption Exponential decay estimates and layer width Delta h ~ mu |ln mu| from prior work [10]
Quoted as 'similar to [10]' in Section 2.4; used in Theorem 2.4 but not re-derived in this paper.
ad hoc to paper Bounded total training error epsilon_app + epsilon_gen + epsilon_opt <= M, and well-definedness of h0, hF, hN
Theorem 3.2 requires this bound for the C2(epsilon_app+epsilon_gen+epsilon_opt) term; it is an external assumption about the non-convex optimization, not proven.
domain assumption Availability of explicit formulas for phi(+/-) and Q0
Algorithm 3.1 must compute these functions; Remark 3.1 concedes this is not available for general A, and all experiments use A=-u.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Deep asymptotic expansion method for solving singularly perturbed time-dependent reaction-advection-diffusion equations." pith.science (2026). https://pith.science/paper/2EVS4EU2

@misc{pith2026250523002,
  author       = {Pith},
  title        = {Pith review of: Deep asymptotic expansion method for solving singularly perturbed time-dependent reaction-advection-diffusion equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2EVS4EU2}},
  note         = {Machine review of arXiv:2505.23002}
}

read the original abstract

Physics-informed neural network (PINN) has shown great potential in solving partial differential equations. However, it faces challenges when dealing with problems involving steep gradients. The solutions to singularly perturbed time-dependent reaction-advection-diffusion equations exhibit internal moving transition layers with sharp gradients, and thus the standard PINN becomes ineffective. In this work, we propose a deep asymptotic expansion (DAE) method, which is inspired by asymptotic analysis and leverages deep learning to approximate the smooth part of the expansion. We first derive the governing equations for transition layers, which are then solved using PINN. Numerical experiments show that the DAE outperforms the standard PINN, gPINN, and PINN with adaptive sampling. We also show its robustness with respect to training point distributions, network architectures, and random seeds.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

[1]

Ainsworth and J

M. Ainsworth and J. Dong,Galerkin neural network approximation of singularly-perturbed elliptic systems, Comput. Methods Appl. Mech. Eng., 402 (2022), p. 115169

work page 2022
[2]

Ainsworth and J

M. Ainsworth and J. Dong,Extended galerkin neural network approximation of singular variational problems with error control, SIAM J. Sci. Comput., 47 (2025), pp. C738–C768

work page 2025
[3]

Arzani, K

A. Arzani, K. W. Cassel, and R. M. D’Souza,Theory-guided physics-informed neural networks for boundary layer problems with singular perturbation, J. Comput. Phys., 473 (2023), p. 111768

work page 2023
[4]

Blum and R

A. Blum and R. Rivest,Training a 3-node neural network is np-complete, in Advances in Neural Information Processing Systems, 1988

work page 1988
[5]

Bottou and O

L. Bottou and O. Bousquet,The tradeoffs of large scale learning, in Advances in neural information processing systems, 2007

work page 2007
[6]

F. Cao, F. Gao, X. Guo, and D. Yuan,Physics-informed neural networks with parameter asymptotic strategy for learning singularly perturbed convection-dominated problem, Comput. Math. Appl., 150 (2023), pp. 229–242

work page 2023
[7]

F. Cao, F. Gao, D. Yuan, and J. Liu,Multistep asymptotic pre-training strategy based on PINNs for solving steep boundary singular perturbation problems, Comput. Methods Appl. Mech. Eng., 431 (2024), p. 117222

work page 2024
[8]

S. Cen, B. Jin, X. Li, and Z. Zhou,Imaging anisotropic conductivity from internal mea- surements with mixed least-squares deep neural networks, J. Comput. Phys., 523 (2025), pp. 113648, 25

work page 2025

Show all 38 references

[9]

Chaikovskii, A

D. Chaikovskii, A. Liubavin, and Y. Zhang,Asymptotic expansion regularization for inverse source problems in two-dimensional singularly perturbed nonlinear parabolic pdes, CSIAM Trans. Appl. Math., 4 (2023), pp. 721–757

2023
[10]

Chaikovskii and Y

D. Chaikovskii and Y. Zhang,Convergence analysis for forward and inverse problems in singularly perturbed time-dependent reaction-advection-diffusion equations, J. Comput. Phys., 470 (2022), p. 111609

2022
[11]

Chaikovskii and Y

D. Chaikovskii and Y. Zhang,Solving forward and inverse problems involving a nonlinear three-dimensional partial differential equation via asymptotic expansions, IMA J. Appl. Math., 88 (2023), pp. 525–557

2023
[12]

K. Cong, G. Li, Y. Sun, and H. Ren,Solving forward problems of boundary-layer problems in ordinary differential equations via physics-informed derivative networks, Phys. Fluids, 37 (2025)

2025
[13]

F ang, S

Z. F ang, S. W ang, and P. Perdikaris,Ensemble learning for physics informed neural networks: a gradient boosting approach, in ICLR 2024 Workshop on AI4DifferentialEquations In Science

2024
[14]

G.-M. Gie, Y. Hong, and C.-Y. Jung,Semi-analytic PINN methods for singularly perturbed boundary value problems, Appl. Anal., 103 (2024), pp. 2554–2571

2024
[15]

L. Guo, H. Wu, X. Yu, and T. Zhou,Monte Carlo fPINNs: Deep learning method for forward 24QIAO ZHU, DMITRII CHAIKOVSKII, BANGTI JIN, AND YE ZHANG and inverse problems involving high dimensional fractional partial differential equations, Comput. Methods Appl. Mech. Eng., 400 (20...

2022
[16]

J. Han, A. Jentzen, and W. E,Solving high-dimensional partial differential equations using deep learning, Proc. Natl. Acad. Sci., 115 (2018), pp. 8505–8510

2018
[17]

Hornik,Approximation capabilities of multilayer feedforward networks, Neural Netw., 4 (1991), pp

K. Hornik,Approximation capabilities of multilayer feedforward networks, Neural Netw., 4 (1991), pp. 251–257

1991
[18]

T. Hu, B. Jin, and Z. Zhou,Solving Poisson problems in polygonal domains with singularity enriched physics informed neural networks, SIAM J. Sci. Comput., 46 (2024), pp. C369–C398, https://doi.org/10.1137/23M1601195

2024 doi
[19]

B. Jin, X. Li, Q. Quan, and Z. Zhou,Conductivity imaging from internal measurements with mixed least-squares deep neural networks, SIAM J. Imaging Sci., 17 (2024), pp. 147–187, https://doi.org/10.1137/23M1562536

2024 doi
[20]

Levashova, A

N. Levashova, A. Sidorova, A. Semina, and M. Ni,A spatio-temporal autowave model of shanghai territory development, Sustainability, 11 (2019), p. 3658

2019
[21]

N. T. Levashova and O. A. Nikolaeva,The heat equation solution near the interface between two media, Model. Anal. Informa. Sist., 24 (2017), pp. 339–352

2017
[22]

L. Lu, X. Meng, Z. Mao, and G. E. Karniadakis,DeepXDE: A deep learning library for solving differential equations, SIAM Rev., 63 (2021), pp. 208–228

2021
[23]

M. D. McKay, R. J. Beckman, and W. J. Conover,A comparison of three methods for select- ing values of input variables in the analysis of output from a computer code, Technometrics, 42 (2000), pp. 55–61

2000
[24]

Nefedov,The existence and asymptotic stability of periodic solutions with an interior layer of burgers type equations with modular advection, Math

N. Nefedov,The existence and asymptotic stability of periodic solutions with an interior layer of burgers type equations with modular advection, Math. Model. Nat. Phenom., 14 (2019), p. 401

2019
[25]

Olchev, K

A. Olchev, K. Radler, A. Sogachev, O. Panferov, and G. Gravenhorst,Application of a three-dimensional model for assessing effects of small clear-cuttings on radiation and soil temperature, Ecol. Modell., 220 (2009), pp. 3046–3056

2009
[26]

Orlov, N

A. Orlov, N. Levashova, and T. Burbaev,The use of asymptotic methods for modelling of the carriers wave functions in the Si/SiGe heterostructures with quantum-confined layers, J. Phys. Conf. Ser., 586 (2015), p. 012003

2015
[27]

Raissi, P

M. Raissi, P. Perdikaris, and G. E. Karniadakis,Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, J. Comput. Phys., 378 (2019), pp. 686–707

2019
[28]

Rudenko,Inhomogeneous Burgers equation with modular nonlinearity: excitation and evolution of high-intensity waves, Dokl

O. Rudenko,Inhomogeneous Burgers equation with modular nonlinearity: excitation and evolution of high-intensity waves, Dokl. Math., 95 (2017), pp. 291–294

2017
[29]

Sidorova, N

A. Sidorova, N. Levashova, and A. Semina,Autowave model of megapolis morphogenesis in the context of inhomogeneous active media, Bull. Russ. Acad. Sci.: Phys., 83 (2019), pp. 91–96

2019
[30]

Sirignano and K

J. Sirignano and K. Spiliopoulos,DGM: A deep learning algorithm for solving partial differential equations, J. Comput. Phys., 375 (2018), pp. 1339–1364

2018
[31]

Sogachev and O

A. Sogachev and O. Panferov,Modification of two-equation models to account for plant drag, Bound.-Layer Meteorol., 121 (2006), pp. 229–266

2006
[32]

K. Tang, X. W an, and C. Yang,DAS-PINNs: A deep adaptive sampling method for solving high-dimensional partial differential equations, J. Comput. Phys., 476 (2023), p. 111868

2023
[33]

W ang, L

L. W ang, L. Zhang, and G. He,Chien-physics-informed neural networks for solving singularly perturbed boundary-layer problems, Appl. Math. Mech., 45 (2024), pp. 1467–1480

2024
[34]

W ang, Y

S. W ang, Y. Teng, and P. Perdikaris,Understanding and mitigating gradient flow pathologies in physics-informed neural networks, SIAM J. Sci. Comput., 43 (2021), pp. A3055–A3081. [35]S. W ang, H. W ang, and P. Perdikaris,On the eigenvector bias of fourier feature networks: Fro...

2021
[36]

W ang, C

Y. W ang, C. Xu, M. Yang, and J. Zhang,Less emphasis on hard regions: curriculum learning of pinns for singularly perturbed convection-diffusion-reaction problems, East Asian J. Appl. Math., (2023)

2023
[37]

J. Yu, L. Lu, X. Meng, and G. E. Karniadakis,Gradient-enhanced physics-informed neural networks for forward and inverse pde problems, Comput. Methods Appl. Mech. Eng., 393 (2022), p. 114823

2022
[38]

Y. Zang, G. Bao, X. Ye, and H. Zhou,Weak adversarial networks for high-dimensional partial differential equations, J. Comput. Phys., 411 (2020), p. 109409

2020
[39]

Zhang and G

L. Zhang and G. He,Multi-scale-matching neural networks for thin plate bending problem, Theor. Appl. Mech. Lett., 14 (2024), p. 100494

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Ainsworth and J

M. Ainsworth and J. Dong,Galerkin neural network approximation of singularly-perturbed elliptic systems, Comput. Methods Appl. Mech. Eng., 402 (2022), p. 115169

work page 2022

[2] [2]

Ainsworth and J

M. Ainsworth and J. Dong,Extended galerkin neural network approximation of singular variational problems with error control, SIAM J. Sci. Comput., 47 (2025), pp. C738–C768

work page 2025

[3] [3]

Arzani, K

A. Arzani, K. W. Cassel, and R. M. D’Souza,Theory-guided physics-informed neural networks for boundary layer problems with singular perturbation, J. Comput. Phys., 473 (2023), p. 111768

work page 2023

[4] [4]

Blum and R

A. Blum and R. Rivest,Training a 3-node neural network is np-complete, in Advances in Neural Information Processing Systems, 1988

work page 1988

[5] [5]

Bottou and O

L. Bottou and O. Bousquet,The tradeoffs of large scale learning, in Advances in neural information processing systems, 2007

work page 2007

[6] [6]

F. Cao, F. Gao, X. Guo, and D. Yuan,Physics-informed neural networks with parameter asymptotic strategy for learning singularly perturbed convection-dominated problem, Comput. Math. Appl., 150 (2023), pp. 229–242

work page 2023

[7] [7]

F. Cao, F. Gao, D. Yuan, and J. Liu,Multistep asymptotic pre-training strategy based on PINNs for solving steep boundary singular perturbation problems, Comput. Methods Appl. Mech. Eng., 431 (2024), p. 117222

work page 2024

[8] [8]

S. Cen, B. Jin, X. Li, and Z. Zhou,Imaging anisotropic conductivity from internal mea- surements with mixed least-squares deep neural networks, J. Comput. Phys., 523 (2025), pp. 113648, 25

work page 2025